Using a Transactor/Revolver Scorecard to Make Credit and Pricing Decisions Mee Chi So Lyn Thomas University of Southampton Hsin-Vonn Seow University of Nottingham Malaysia Campus
The Standard Approach All credit card applicants Use a credit scorecard to make credit decision Accepted credit card users Transactors Revolvers
Transactors and Revolvers Transactors all are Goods profit from merchant service charge only Revolvers higher chance to be Bads main profit from interest on the balance Important in terms of default risk More important in terms of profitability Could we estimate this when making credit decision?
Our Proposed Approach All credit card applicants Use a credit scorecard Use a tran/rev scorecard A combined score Accepted credit card users
A Combined Score Define: T-Transactor, R-Revolver, G-Good, B-Bad A score gives the probability the new customer is likely to be Good: P G x = P T x P(G x, T) + P R x P G x, R Since no Transactor can default, P(G x, T)=1. Therefore, P(G x) = P T x + P R x P G x, R
To Develop The Two Scorecards The Tran/Rev Scorecard Using all data s t x = ln P T x P T x = 1, P R x = 1 P R x 1+e s t (x) 1+e s t (x) The Good/Bad Scorecard to Revolvers Only Using revolvers data only s R x = ln P(G x,r) P(B x,r) P G x, R = P B x, R = 1, 1+e s R (x) 1 1 + e s R(x)
A Numeric Example Credit card data from a Hong Kong bank Accounts opened 2002-2005; Outcome period: 2006 Total 6,308 accounts: 1,577 Bad, 4,731 Good List of variables: Occupation, Education type, Citizenship, Residential type, Employment status, Annual income, Months with bank and Age Use weight-of-evidence for all characteristics Use stepwise logistic regression Use ten-fold cross validation Good Bad Total Transactor(count and column %) 2958 (63%) 0 (0%) 2958 (47%) Revolver(count and column %) 1773 (37%) 1577 (100%) 3350 (53%) Total(count) 4731 1577 6308
Coefficients for All Scorecards Variable (WoE) Standard Scorecard (Event=Good) Coefficient (Mean) Coefficient (S.D.) Transactor/Revolver Scorecard (Event=Transactor) Coefficient (Mean) Coefficient (S.D.) Good/Bad Scorecard (by Revolvers only) (Event=Good) Coefficient (Mean) Coefficient (S.D.) Intercept 1.3929*** 0.0072-0.1254** 0.0053 0.4291*** 0.0125 Occupation 0.7796*** 0.0292 0.5994*** 0.0209 0.5287*** 0.0328 Education type 1.3961*** 0.0678 0.4701** 0.0556 1.4056*** 0.0745 Citizenship 1.2286*** 0.0743 0.9286*** 0.0255 0.8748** 0.0850 Residential type Employment status Months with bank Annual income 1.1147*** 0.0578 0.1146**# 0.1848 1.1741*** 0.0187 - - 0.6864*** 0.0391 0.3951** 0.0430 0.7998*** 0.0209 0.2150** 0.0290 0.7486*** 0.0707 - - 0.8240*** 0.0259 - - Age - - 0.2660** 0.0291 - - *** significant at 0.0001; ** significant at 0.05; #selected by three models only. For models do not pick up the variable, we assume the coefficients equal 0.
Gini Coefficients Gini coefficients for different scorecards Standard Tran/Rev G/B with Rev only Combined ROC Contrast Test Results between Standard and Combined Validation 1 0.592 0.45 0.596 0.592 0.039(0.8434) Validation 2 0.47 0.394 0.47 0.474 0.5008(0.4791) Validation 3 0.556 0.45 0.542 0.558 0.0488(0.8251) Validation 4 0.52 0.42 0.518 0.511 2.1165(0.1457) Validation 5 0.482 0.426 0.486 0.48 0.0789(0.7788) Validation 6 0.526 0.4 0.53 0.53 0.3427(0.5583) Validation 7 0.512 0.428 0.506 0.508 0.9092(0.3403) Validation 8 0.524 0.44 0.522 0.532 0.057(0.8114) Validation 9 0.548 0.432 0.54 0.542 0.8449(0.358) Validation 10 0.484 0.47 0.48 0.49 1.2506(0.2634) Average 0.522 0.431 0.519 0.522
ROC Curves for Two Folds Validation 1 Validation 4
Credit Card Profitability Model If the interest rate offered on credit cards is r, the corresponding expected monthly profit for the lender is: e r, p : the expected monthly profit of a customer with hazard rate p 1 E(r) = e r, p p (r) q r, p f p dp f(p): The distribution of the Good hazard rates p (r): the cut-off level of the hazard rate of being Good q(r, p): a customer s take-up probability
Risk and Take Function Population s hazard risk distribution F p = 0, p < 0.5 2p 2 2p + 0.5,0.5 p < 1 1, p = 1 Take Function (Phillips, 2005; Thomas, 2009) q r, p = Max{0, 3 10r 2p} For example, if interest rate r = 3% and the hazard rate is 0.9, the take rate is q 3%, 0.9 = 3 10 3% 2(0.9) = 90%
The Expected Profit: e(r, p) Income interchange fees and interest on balance The expected profit: e r, p = Interchange fee Average Purchase +P(Not Default in N period) (Repayment in N period) +P Default in N period Recovery via Collection Given an interest rate, find the optimal cut-off probability by e r, p = 0 p = (1 m)(1 + r F) N l D (1 + r) N 1 + l D 1 l D Parameters required m Interchange rate r Interest per period of the credit cards r F Interest rate at which lenders can borrow money per period l D Loss given default (LGD) N Average number of periods before repaying the average purchase 1/N
To Estimate N N: average number of periods before a purchase is paid off B: average balance carried over per period per customer P: average amount purchased per period per customer C: average repayment amount per period per customer Interest paid + Ave. Expenditure = Ave. Repayment, i.e. rb + P = C.. (*)
To Estimate N (cont.) B Br P t t+1 t+n C C C Assume the user pays off the costs in the order they are incurred 1 + r B + P = CN.. ( ) Using (*) and (**), B = C P r N = 1 + r B + P C = B + C C
A Numerical Example m Interchange rate 2% r F Interest rate at which lender can borrow money per period 1% l D LGD 60% P Average purchase per period 51 C Average repayment per period 60 Using the above parameters and the equations listed before, we can find the expected profit and the corresponding hazard rate: r p (p ) 12 E(r) 3% 0.969 0.687 2.084 2% 0.983 0.817 1.783 4% 0.957 0.590 2.040
With a Tran/Rev Scorecard If the interest rate offered on credit cards is r, the corresponding expected monthly profit for the lender is: E(r) = dt 1 p R (t) e(r, p) q r, p f p, t dp
Risk and Take Function Same take function q r, p = Max{0, 3 10r 2p} Population s hazard risk distribution 0 p < 0.5 2p 2 2p + 0.5 0.5 p < 1,2p 1 > t F p, t = 2tp t 0.5t 2 0.5 p < 1, 2p 1 t 1 p = 1, t = 1 Percentage of transactors 1 t = tdt f(p, t) dp = 2/3 0 0 1
The Expected Profit: e(p R, t) t: the transactor score p R : the Good hazard rate from the Revolver Good/Bad scorecard e p R, t = t (Interchange fee) + 1 t [Interchange fee Average Purchase + P(Not Default in N period) (Repayment in N period) +P Default in N period Recovery via Collection ] Given an interest rate, find the optimal cut-off probability by e p R, t = 0 p R = (1 + r F ) l D (1 + r) N R 1 tp T 1 m 1 + 1 m + l D 1 (1 t)p T 1 + r F l D 1/N R
With a Tran/Rev Scorecard N: average number of periods before a purchase is paid off B: average balance carried over per period per customer P: average amount purchased per period per customer C: average repayment amount per period per customer Revolvers Transactors All users rb R + P R = C R 1 + r B R + P R = C R N R B T = 0 C T = P T N T = 1 P = P T P T + 1 P T P R, C = P T C T + 1 P T C R, B = 1 P T B R, N = αn R + (1 α) where α = C R (1 P T ) C T P T +C R (1 P T )
The Numerical Example with Tran/Rev Split P Average purchase per period 51 C Average repayment per period 60 P(T) Percentage of transactors 2/3 P T P for Transactor 72 C T C for Transactor 72 P R P for Revolver 9 C R C for Revolver 36 p R (t): the optimal cut-off corresponding to transactor score t r t E(r) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 3% 0.960 0.959 0.957 0.956 0.953 0.950 0.943 0.929 0.839 0.000 0.000 0.330 4% 0.924 0.922 0.919 0.915 0.909 0.898 0.872 0.000 0.000 0.000 0.000 0.399 2% 0.982 0.981 0.980 0.979 0.977 0.975 0.971 0.963 0.940 0.000 0.000 0.311
Conclusion & Possible Extensions Build a scorecard to estimate P(T x) How to use the score in profitability modelling The model with Tran/Rev: acknowledge the profitability of Transactors so that the estimation on profitability is more accurate and offer a different price Use it for Churn prediction?
Thank you